Endocrine Abstracts (2017) 49 EP1377 | DOI: 10.1530/endoabs.49.EP1377

Predictive value of CHAID Algorithm in the diagnosis of malignancy in thyroid nodules with Bethesda III (AUS/FLUS) cytology

Reyhan Ersoy1, Cevdet Aydin1, Oya Topaloglu1, Didem Ozdemir1, Husniye Baser2, Fatma Neslihan Cuhaci1, Abbas Ali Tam1, Berna Ogmen2, Sefika Burcak Polat1, Fatma Dilek Dellal2, Ahmet Dirikoc1, Nagihan Bestepe2, Mehmet Kilic3, Aysegul Aksoy Altinboga4 & Bekir Cakir1


1Department of Endocrinology and Metabolism, School of Medicine, Ankara Yildirim Beyazıt University, Ankara, Turkey; 2Department of Endocrinology and Metabolism, Ankara Ataturk Education and Research Hospital, Ankara, Turkey; 3Department of General Surgery, School of Medicine, Ankara Yildirim Beyazıt University, Ankara, Turkey; 4Department of Pathology, Ankara Ataturk Education and Research Hospital, Ankara, Turkey.


Introduction: ‘Artificial intelligence’ is an information-processing technology developed and inspired by the information processing technique of human brain. Artificial intelligence applications are used with an increasing ratio in medicine; particularly in the fields of breast cancer, radiology, cardiology, health management and drug effects analysis. There is not any study concerning thyroid diseases and artificial intelligence applications in the literature. In this study, we aimed to use an artificial intelligence application - CHAID (Chi-Squaered Automatic Interaction Detection) algorithm- to predict malignancy risk in thyroid nodules with Bethesda III –atypia of undetermined significance/follicular lesion of undetermined significance (AUS/FLUS)- cytology.

Methods: Data of 3300 patients (6863 nodules) evaluated by our multidisciplinary council and operated between 2005–2016 were analysed retrospectively. There were 364 patients with 385 nodules with AUS/FLUS cytology in. 65 parameters including clinical, laboratory, ultrasonographical, cytological and histopathological features of each nodule were evaluated. The diagnostic value of CHAID algorithm was determined considering histopathological diagnosis as the reference method.

Results: There were 75 male (20.6%) and 289 female (79.4%) patients. Histopathological diagnosis was benign in 282 (73.2%) and malignant in 103 (26.8%) nodules. Analysis with CHAID algorithm revealed that presence of thyroiditis/chronic thyroiditis ultrasonographically, and presence of nuclear groove and intranuclear inclusions cytologically were predictive for malignancy. This algorithm had a sensitivity of 49.5%, specifity of 96.8%, positive predictive value of 85%, negative predictive value of 16% and accuracy of 84.15%.

Conclusion: The main issue in thyroid nodules is to differentiate benign and malignant lesions. While doing this, clinician should try to avoid unnecessary ultrasonography examinations, fine needle aspiration biopsy and surgical approach. Superfluous interventions cause increased economical burden for both the patient and the country. Use of artificial intelligence applications in clinical practice might help to reduce unnecessary diagnostic procedures and surgical interventions in thyroid nodules with AUS/FLUS. cytology.